distributed systems

This paper addresses the lack of modular communication modeling and insufficient reliability analysis in distributed consensus networks. We propose the first generic communication abstraction framework for crash- and Byzantine-fault-tolerant protocols, unifying core communication components across RAFT, Paxos, PBFT, and HotStuff. Innovatively, we develop a probabilistic graphical model-based method to quantify consensus reliability, enabling joint evaluation of failure probability and latency under link loss and node failures. We further design two protocol-level latency optimization mechanisms and implement a failure-rate-controllable communication layer. Theoretical analysis is empirically validated on a RAFT prototype: end-to-end latency is significantly reduced. Our framework provides a reusable modeling and design foundation for low-failure, low-latency consensus systems.

Existing Raft reconfiguration schemes rely on a centralized coordinator and require full-cluster downtime, introducing single-point-of-failure vulnerabilities and correctness risks. This paper proposes ReCraft: a coordinator-free dynamic reconfiguration mechanism supporting split/merge operations and fine-grained membership changes. Its core is a self-contained, multi-level reconfiguration protocol, realized through extensions to the Raft state machine and a redesigned distributed consensus logic—formally verified for safety and liveness using TLA+. Implemented in etcd, ReCraft demonstrates that reconfiguration blocks only necessary log submissions, incurs <8% throughput degradation, reduces split/merge latency by 57%, and completely eliminates single-point failures inherent in centralized coordination.

Standard Raft suffers from high message round-trips, strong leader dependency, and substantial communication overhead in dynamic distributed environments. To address these limitations, this paper proposes Fast Raft—a hierarchical consensus protocol. Its core innovations include a layered architecture and a fast path for log commitment, enabling significant reductions in message rounds and leader load while preserving Raft’s safety and liveness guarantees. Node communication is implemented via gRPC, and the system is deployed across multiple AWS Availability Zones using Kubernetes to ensure high availability and scalability for empirical evaluation. Experimental results demonstrate that, under low packet-loss conditions, Fast Raft achieves a 42% throughput improvement and reduces average log-commit latency by 37%, substantially enhancing consensus performance and practicality in dynamic settings.

Under asynchronous network partitions, the CAP theorem imposes a fundamental trade-off between consistency and availability. This paper challenges the conventional binary assumption—requiring either strong consistency or global availability—by proposing the “partial progress” conjecture: permitting responsive service and non-zero throughput for a subset of clients during partitions. Method: We formally define “partial progress” and prove its theoretical feasibility in the asynchronous model. We design CASSANDRA, a novel consensus protocol that leverages causal ordering and vector clocks to enable locally ordered replica processing without global synchronization, augmented by a lightweight coordination mechanism to ensure local consistency. Results: Experimental evaluation under simulated partition scenarios demonstrates that CASSANDRA sustains service availability for over 85% of clients, achieves linear throughput scaling, and significantly outperforms mainstream protocols such as Paxos and Raft.

Traditional consensus protocols rely on deterministic f-threshold fault models that struggle to capture the complex failure behaviors observed in real-world systems, thereby limiting optimization of performance and cost. This work proposes a novel consensus mechanism grounded in a probabilistic fault model, which incorporates machine-level failure curves and abandons the rigid majority quorum constraint in favor of dynamic, non-traditional quorum strategies. By more accurately reflecting actual operating conditions, the proposed approach substantially enhances system reliability, efficiency, cost-effectiveness, and sustainability.

Existing blockchain nodes struggle to meet the demands of high throughput and high reliability due to limitations in scalability, availability, and fault tolerance. This work proposes a distributed node architecture based on the RAFT consensus protocol, which employs a leader-follower coordination mechanism to centralize stateless operations at the leader while replicating stateful operations across the cluster. By integrating a state partitioning strategy with concurrent Merkle tree update techniques, the design achieves load balancing and strong consistency. The proposed approach significantly enhances system scalability, fault tolerance, and overall performance, while effectively reducing the overhead associated with smart contract execution and Merkle tree maintenance.

This work addresses the tight coupling between reconfiguration mechanisms and consensus protocols in existing state machine replication (SMR) systems, which hinders independent component upgrades and incurs significant downtime. To overcome this limitation, the authors propose Gauss, a reconfiguration engine that decouples the internal consensus log from an externally exposed clean log through a dual-log architecture. This separation enables modular evolution of membership changes, fault-tolerance thresholds, and even the underlying consensus protocol itself. By isolating reconfiguration logic from consensus execution, Gauss substantially reduces system maintenance complexity and supports seamless, independent upgrades of individual components. Evaluation on the Rialto blockchain platform demonstrates that Gauss facilitates near-zero-downtime transitions between different consensus protocols, achieving highly available and adaptable SMR deployments.

This work addresses the challenges of efficiently disseminating and storing erasure-coded data in distributed systems, where unpredictable network latency, complex recovery procedures, and reliance on centralized coordination hinder performance and scalability. To overcome these limitations, the paper proposes Rafture, a novel distribution algorithm that integrates the Raft consensus protocol with erasure coding. Rafture introduces a post-dissemination pruning mechanism to dynamically adjust storage overhead, employs a fixed-threshold erasure code combined with a two-dimensional chunk allocation strategy to simplify recovery without requiring additional metadata, and enables nodes to make autonomous decisions based solely on local information, thereby eliminating dependence on a centralized leader. Experimental results demonstrate that Rafture significantly reduces long-term storage costs while improving recovery efficiency in highly dynamic network environments.

This work proposes a decentralized multi-agent collaborative scheduling framework to address the limitations of centralized schedulers in distributed scientific workflows, including single points of failure, poor scalability, and weak dynamic adaptability. The framework employs a hierarchical consensus architecture supporting up to a thousand agents, integrating a highly fault-tolerant task management mechanism with a data-aware scheduling strategy that jointly considers resource states and data locality. Experimental evaluations on the FABRIC testbed demonstrate that the system scales to 1,000 agents, achieves over 99% task completion rate under single-point failures, incurs no more than a 7.5% performance degradation when 50% of nodes fail, and reduces task scheduling latency by 97–98% compared to baseline approaches.